Extremely exciting results and analysis.
Introducing our work: RoboDawn 🌅 A VLM controls robots with zero robot training.
📈 In-context learning: one demo lifts RoboTwin with GPT-6-Astra 2.0 53.2% → 73.6%, RoboDojo 35.67% → 47.17%
🔁 Test-time scaling: 31.2% → 47.2% with more commands
🔓 710 episodes and sessions
00:00
